IC-1321Vicuna-13B, instruction-tuned from LLaMA-13B on user conversations, exhibits worse calibration than its base model LLaMA-13B

Xin Liu, Muhammad Khalifa, Lu Wang

SourceLitCab: Lightweight Language Model Calibration over Short- and Long-form Responses

The paper compares the calibration of Vicuna-13B (v1.3), which was fine-tuned from LLaMA-13B on user-shared conversations, against LLaMA-13B itself. Vicuna-13B shows higher ECE and Brier scores across all tasks, indicating that the instruction-tuning stage degraded calibration. For example, on NQ, Vicuna-13B has ECE 0.204 versus LLaMA-13B at 0.123. The authors conclude that additional fine-tuning with limited-purpose data can diminish a model's calibration.

Evidence
correlational
Key metric
NQ ECE: LLaMA-13b 0.123, Vicuna-13b 0.204; TruthfulQA ECE: LLaMA-13b 0.121, Vicuna-13b 0.200; NQ Brier: LLaMA-13b 0.180, Vicuna-13b 0.224
Caveat
The comparison is between two specific checkpoints; the paper does not isolate whether the degradation is due to the fine-tuning data, the fine-tuning procedure, or both.
Model
Vicuna, LLaMA
Datasets
Natural Questions / NaturalQA [eval], TruthfulQA / TruthfulQA MC1 [eval]
Methods
Expected Calibration Error / Integral Calibration Error (ECE) [eval], Brier score [eval]
Related findings
IC-1319, IC-1320
Extraction
automatic-extraction